A Unified Lightweight Attention Module for Road Classification and Crack Segmentation

Automated health maintenance of transport infrastructure is pivotal for ensuring public safety and operational efficiency within intelligent transportation systems. While deep learning has revolutionized surface distress diagnosis, a critical trade-off persists between diagnostic accuracy and computational efficiency, particularly for deployment on resource-constrained edge devices. To bridge this gap, we propose the anchor gated self-attention (AGSA), a lightweight plug-and-play module. AGSA introduces sparse learnable anchors as semantic hubs to decouple global dependency modeling, reducing computational complexity from quadratic to linear. Furthermore, it incorporates a novel anchor-guided spatial gating mechanism that dynamically fuses global semantics with local spatial details, effectively highlighting fault regions while suppressing environmental noise. Extensive experiments demonstrate the superiority of AGSA across diverse tasks. In road surface classification, AGSA consistently enhances performance across multiple backbones on the large-scale road surface classification dataset benchmark and a challenging self-collected winter dataset, achieving state-of-the-art accuracy in extreme conditions such as ice and snow coverage. Moreover, in dense prediction tasks, AGSA significantly improves crack segmentation precision on the Crack500 dataset, yielding notable gains in IoU and Dice scores for subtle fracture detection. These results confirm AGSA as a scalable, high-performance solution for real-time, automated infrastructure inspection.